A state-of-the-art article was published in the KI Wissen project. The article covers approaches, methods, and references from the three core areas of knowledge integration, extraction, and conformity and thus shows a far-reaching state of knowledge that provides important indications and directions for further work in the project.
For the article with the title "Knowledge Augmented Machine Learning with Applications in Autonomous Driving: A Survey", a total of 46 authors from the project compiled over 885 references. The article thus provides a comprehensive insight into previous research on the topic of knowledge and AI.
The article starts with a brief introduction of the use cases pursued in the project: perception, situation recognition, and planning. These three use cases are referenced again in the respective subsections on the approaches and methods. This way, in addition to a general presentation of the approaches, a direct connection of the presented methods to autonomous driving is established.
Following the introduction and use cases, possible forms of knowledge representation such as knowledge graphs and embeddings are explained. This is followed by approaches to integration, e.g. extension of the cost function using regularizers and constraints, neuro-symbolic integration, data augmentation and data fusion. Approaches to knowledge transfer, e.g. from pre-trained models or by means of active learning, conclude the area of knowledge integration.
In the area of extraction, the focus is on symbolic approaches, primarily rule extraction methods and methods that provide output based on natural language. Furthermore the article considers approaches that provide visual representations (e.g., heat maps) for explaining and tracking decisions. The article concludes with methods for knowledge conformance, in particular existing methods for estimating uncertainties, and determining causal relationships.
The article can be accessed via arXiv. During the course of the project it is planned to revise and create new versions of the article in order to ensure an up-to-date overview of research in this field.
Image: unsplash/Patrick Tomasso